data attribution for robot learning
Find the demonstrations that are hurting your policy.
skyu scores every episode in a robot-learning dataset by how much it surprises your trained policy, flags the suspect tail, and writes a reproducible report. It runs offline, on your hardware.
Placeholder data — 12 injected corruptions, all recovered in the flagged tail.
How it works
Point
Give skyu a LeRobot-format dataset and a policy checkpoint. Nothing else — no labels, no cloud account, no instrumentation of your training run.
skyu score --dataset D --checkpoint C
Score
Every episode gets a surprise score: mean training loss over K seeded draws from your policy. Episodes the policy can't explain rise to the top.
surprise = mean loss over K draws
Review
The report flags the suspect tail — top-N% plus elbow detection — and every number in it carries the exact command that reproduces it.
→ report.html · single file
Blind validation
Figures marked TK are withheld pending approval — nothing here is invented.of injected corruptions recovered in the flagged tail, scored blind against held-out corruption labels.
more bad episodes found per hour of review than a random audit of the same budget.
rank AUC across corruption types — sensor dropout, label swaps, truncated episodes.
Every result ships with the seed, the checkpoint hash, and the command to rerun it. skyu backtest reproduces the full table on your machine.
Built for teams that can't upload their data
Offline-first
Runs entirely on your hardware. No telemetry, no uploads — your demonstrations never leave your machines.
Deterministic
Seeded draws, pinned kernels. Run it twice and the reports are bit-identical.
One-file report
The output is a single dependency-free HTML file. Archive it, diff it, attach it to a PR.
Evaluation license
A plain evaluation license for design partners. No procurement maze to try it.
See it on your own dataset.
We're onboarding a small number of design partners. Tell us where to reach you — a work email is enough.
No pricing call, no newsletter. We reply to robotics teams with real datasets.